A selective deep stacked denoising autoencoders ensemble with negative correlation learning for gearbox fault diagnosis

Abstract

Vibration signals are widely used as an effective way to fulfill gearbox fault diagnosis. However, it is quite challenging to extract effective fault features from noisy vibration signals and then to construct a reliable fault diagnosis model. This paper proposes a selective stacked denoising autoencoders (SDAE) with negative correlation learning (NCL) (SSDAE-NCL) for gearbox fault diagnosis. The component SDAEs are firstly constructed to extract effective fault features from vibration signals in the unsupervised-learning phase of SSDAE-NCL. Based on the extracted features, NCL is used to fine-tune the SDAE components to construct component classifiers in the supervised-learning phase of SSDAE-NCL. Finally, a selective ensemble is finished based on these divers and accurate component SDAEs for gearbox fault diagnosis. The motivation for developing ensemble of deep neural networks (DNNs) is that they can achieve higher accuracy and applicability than single component in machinery fault diagnosis. Furthermore, it can make an overall ensemble model easy to use in real cases for users, because it does not need too much prior knowledge about setup of a DNN model. The effectiveness of this SSDAE-NCL-based fault diagnosis method has been verified by experimental results on the vibration signal data from a gearbox test rig. The results illustrate that SSDAE-NCL learns effective discriminative features from vibration signals and achieves the better diagnosis accuracy in comparison with those typical DNNs (e.g., SDAE, deep belief network (DBN)).